Deep Feature Flow -CVPR 2017

Video Detection and Segmentation

这是一篇很有想法(novel)的工作,尤其是在video领域,作者是来自MSRA的Xizhou Zhu 和DaiJifeng,他们的一系列工作我将在本博客中给予更多介绍,这次介绍的是他们的工作:deep feature flow

Deep Feature Flow

Deep feature flow 正如其名,它是用了deep feature (appearance 信息)和 光流(motion 信息)来对视频中的每一帧做建模,它的核心思想是利用指定的关键帧(key frame)和其他帧(current frame)到关键帧的光流信息,把关键帧送入到deep CNN中得到deep feature,而对于其他帧算出与关键帧的光流,再用光流把关键帧的deep feature propagate (相加,在做双线性差值)到当前帧,这时候就可大大减少计算量,因为对于非关键帧的话就不用再去过CNN了。对于新得到的特征,接下来就可对其进行不同任务处理,比如分割或者检测。整个思路如下图:


首先看它的Motivation:


作者是通过观察resnet-101 模型 deep feature,发现通过光流把前一帧warp到后一帧,发现和用resnet-101 直接inference很接近,于是就想到这种方法。

Inference:

首先给出它的formal notation(ps:就是为了更加正式,其实想法很简单)



N 代表后续任务网络,分割或者检测,取决于当前的task,N 代表提取特征的网络。

重点是第11行,做propagation。

c代表channel, 它是对每个channel做相同的差值(双线性差值,为啥用,是因为这种差值方式是可微的,为了后续的end to end trainning)操作,G代表双线性差值核函数,S代表scale 是用来对key frame feature 和 current frame feature做scale 操作。最后整个式子可以写成 f = W(f, M,S), M代表warp后的feature。

Training

考虑一种情况,为啥不用类似与在video action recognition 任务中方法先把光流计算出来直接做warp呢?作者认为这种作为不适合recognition task,会有很大的variation,个人理解是分割和检测任务不适合这么做,因为相对action recognition,分割和检测的任务是更细粒度的场景理解和目标理解,所以作者使用了flow net作为光流计算, flow net是可以学习,可以自适应不同场景的运动物体建模。他采用的最简单的flow net版本(flow net-S)。

接着作者分析了加速时间,就是可以理解为,warp的时间要远远小于用deep CNN做inference的时间。

不足

1,很明显一点就是key frame的选择,作者这里用了固定的策略,所以今年CVPR2018有很多文章在key frame选择上提出新方法。

2,feature representation的不足,由于flow warp后的feature会变弱,所以对于非关键帧的地方,效果肯定比单帧做检测或分割效果要差。

实验结果

作者进行了2个任务的实验:

1,Cityscapes (分割)

2,ImageNet VID (视频目标检测)

结果如下:


其中还对比了如下不同的架构: (SFF 算法->一种计算光流的方法,这样可以把光流提前计算好)


代码:

https://github.com/msracver/Deep-Feature-Flow

第一次写论文笔记:多多包涵,不足之处请指出~~


Key Features Learn advanced techniques in deep learning with this example-rich guide on Google's brainchildExplore various neural networks with the help of this comprehensive guideAdvanced guide on machine learning techniques, in particular TensorFlow for deep learning. Book Description Deep learning is the next step after machine learning. It is machine learning but with a more advanced implementation. As machine learning is no longer an academic topic, but a mainstream practice, deep learning has taken a front seat. With deep learning being used by many data scientists, deeper neural networks are evaluated for accurate results. Data scientists want to explore data abstraction layers and this book will be their guide on this journey. This book evaluates common, and not so common, deep neural networks and shows how these can be exploited in the real world with complex raw data using TensorFlow. The book will take you through an understanding of the current machine learning landscape then delve into TensorFlow and how to use it by considering various data sets and use cases. Throughout the chapters, you'll learn how to implement various deep learning algorithms for your machine learning systems and integrate them into your product offerings such as search, image recognition, and language processing. Additionally, we'll examine its performance by optimizing it with respect to its various parameters, comparing it against benchmarks along with teaching machines to learn from the information and determine the ideal behavior within a specific context, in order to maximize its performance. After finishing the book, you will be familiar with machine learning techniques, in particular TensorFlow for deep learning, and will be ready to apply some of your knowledge in a real project either in a research or commercial setting. What you will learn Provide an overview of the machine learning landscapeLook at the historical development and progress of deep learningDescribe TensorFlow and become very familiar with it both in theory and in practiceAccess public datasets and use TF to load, process, clean, and transform dataUse TensorFlow on real-world data sets including images and textGet familiar with TensorFlow by applying it in various hands on exercises using the command lineEvaluate the performance of your deep learning modelsQuickly teach machines to learn from data by exploring reinforcement learning techniques.Understand how this technology is being used in the real world by exploring active areas of deep learning research and application.
评论 9
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值